On August 11, Nvidia Laid Out a Plan to Push 2 Megawatts Down a Single Rack Row

Nvidia will release an 800 volt DC power rack in the second half of this year. It bolts onto the alternating current wiring a building already has, without tearing out the electrical plant.

Why scaling AI compute performance requires a new power architecture (Nvidia blog, 2026-08-11)

출처: blogs.nvidia.com

The roadmap circulated first through Taiwanese press, but the source is Nvidia itself. The same plan went up on the company blog on August 11.

Why raise the voltage

Inside a rack today, chips are fed at 54 volts DC. Low voltage means high current for the same power, and high current means thick copper. Put simply, pushing the same water through a thin pipe takes pressure, and the pipe heats up. Raise the pressure and a thinner pipe carries the same flow. By Nvidia's own count, the busbars in a single one megawatt rack can hold up to 200 kilograms of copper. Scale that to a gigawatt and it is 200,000 kilograms. The company says the higher voltage cuts copper use by 45% and improves end-to-end delivery efficiency by up to 5% against the 54 volt approach.

Nvidia 800 VDC architecture will power the next generation of AI factories (Nvidia developer blog, 2025-05-20)

출처: developer.nvidia.com

Buildings already running come first

What matters about the second half product is not performance, it is the retrofit threshold. Because the conversion happens inside the rack row rather than at the building switchgear, a data center finished last year can swap racks and move on. The next step lands in 2027. An overhead DC busway pushes up to 2 megawatts into an entire rack row. After that comes facility-level conversion of grid power straight to DC, and that part is a story about new construction.

Why Google and Microsoft are standing alongside

Nvidia did not write this specification alone. It published through the Open Compute Project with Google and Microsoft, and says more than 80 equipment makers are already building to it. Keeping the spec open reads less like a plan to sell power gear and more like a plan to remove a constraint. The money spreads instead to power silicon vendors such as Infineon, Navitas and Texas Instruments, and to distribution equipment makers such as Vertiv, Schneider Electric and Eaton. For an investor the hook here is not Nvidia's next GPU. It is who gets to sell the parts needed to plug that GPU in. If a row power center is running in a real data center inside 2027, this specification has become the standard. If 2027 passes and the rack-level hybrid is still the state of the art, the building side never caught up.